Randomized Composable Core-sets for Distributed Optimization Vahab Mirrokni

Transcription

1 Randomized Composable Core-sets for Distributed Optimization Vahab Mirrokni Mainly based on joint work with: Algorithms Research Group, Google Research, New York Hossein Bateni, Aditya Bhaskara, Hossein Esfandiari, Silvio Lattanzi, Morteza Zadimoghaddam

2 Our team: Google NYC Algorithms Research Teams Market Algorithms/ Ads Optimization (search & display) Tools: e.g. Clustering common expertise: online allocation problems Large-Scale Graph Mining/Distributed Optimization Infrastructure and Large-Scale Optimization Tools: e.g. Balanced Partitioning

3 Three most popular techniques applied in our tools Local Algorithms: Message Passing/Label Propagation/Local Random Walks e.g., similarity ranking via PPR etc, Connected Components Connected components code that s times faster the state-of-the-art Embedding/Hashing/Sketching Techniques e.g., linear embedding for balanced graph partitioning to minimize cut Improves the state-of-the-art by 26%. Improved flash bandwidth for search backend by 25%. Paper appeared in WSDM 16. Randomized Composable Core-sets for Distributed Computation: This Talk

4 Proprietary + Confidential Agenda Composable core-sets: Definitions & Applications Applications in Distributed & Streaming settings Applications: Feature Selection, Diversity in Search & Recom. Composable Core-sets for Four Problems: Survey Diversity Maximization(PODS 14, AAAI 17), Clustering(NIPS 14), Submodular Maximization(STOC 15), and Column Subset Selection (ICML 16) Sketching for Coverage Problems (on arxiv) Sketching Technique

5 Composable Core-Sets for Distributed Optimization Run ALG in each machine Machine 1 T1 T1 T2 Machine 2 T2 S1 Run ALG on selected items to find the final output set S2 Selected Items Input Set Tm Output Set Sm Machine m Tm

6 Composable Core-sets Setup: Consider partitioning data set T of elements into m sets (T1,T2,...,Tm). Goal: Given a set function f, find a subset S* with Find: small core-set,,.,, optimizing f(s*). such that optimum solution in union of core-sets approximates the optimum solution of T

7 Application in MapReduce/Distributed Computation Run ALG in each machine Machine 1 T1 T1 T2 Machine 2 T2 S1 Run ALG on selected items to find the final output set S2 Selected Items Input Set Tm Output Set Sm Machine m Tm E.g., two rounds of MapReduce

8 Application in Streaming Computation Streaming Computation: Processing sequence of n data points on the fly Limited storage Use C-composable core-set of size k, for example: Chunks of size, thus number of chunks is Compute core-set of size k for each chunk Total space:

9 Overview of recent theoretical results Need to solve (combinatorial) optimization problems on large data 1. Diversity Maximization, PODS 14 by IndykMahdianMahabadiMirrokni for Feature Selection in AAAI 17 by AbbasiGhadiriMirrokniZadimoghaddam 2. Capacitated ℓp Clustering, NIPS 14 by BateniBhaskaraLattanziMirrokni Submodular Maximization, STOC 15 by MirrokniZadimoghaddam Column Subset Selection (Feature Selection), ICML 16 by Alschulter et al. Coverage Problems: Submitted by BateniEsfandiariMirrokni

10 Applications: Diversity & Submodular Maximization Diverse suggestions Play apps Campaign keywords Search results News articles YouTube videos Data summarization Feature selection Exemplar sampling

11 Feature selection We have Data points (docs, web pages, etc.) Features (topics, etc.) Goal: pick a small set of representative features Emotion Hotel Finance Hospital Money Business Weather Cloud Laundry Movie Gaming Smartphone Car World Theatre Software Laptop Boat Home Camera Education Security Train Shopping

12 Five Problems Considered General: Find a set S of k items & maximize/minimize f(s). Diversity Maximization: Find a set S of k points, and maximize the sum of pairwise distances i.e. max diversity(s) =. Capacitated/Balanced Clustering: Find a set S of k centers and cluster nodes around them while minimizing the sum of distances to S. Coverage/Submodular Maximization: Find a set S of k items. Maximize submodular function f(s). Generalizing set cover. Column subset selection: Given a matrix A, find a set S of k columns. Minimize

13 Diversity Maximization Problem Given: A set of n points in a metric space (X, dist) Find a set S of k points Goal: maximize diversity(s) i.e. diversity(s) = sum of pairwise distances of points in S. Background: Max Dispersion (Halldorson et al, Abbassi et al) Useful for feature selection, diverse candidate selection in Search, representative centers...

14 Core-sets for Diversity Maximization Two rounds of MapReduce Run LocalSearch on each machine Machine 1 T1 T1 Input Set T2 Machine 2 T2 S1 Run LocalSearch on selected items to find the final output set S2 Selected Items Output Set Sm Tm Machine m Tm Arbitrary Partitioning works. Random partitioning is better.

15 Composable Core-set Results for Diversity Maximization Theorem(IndykMahabadiMahdianM. 14): The local search algorithm computes a constant-factor composable core-set for maximizing sum of pairwise distances in 2 rounds: Theorem(EpastoM.ZadiMoghaddam 16): A sampling+greedy algorithm computes a randomized 2-approximate composable small-size core-set for diversity maximization in one round. randomized: works under random partitioning small-size: size of core-set is less than k.

16 Distributed Clustering Problems Clustering: Divide data into groups containing nearby points Minimize: k-center : k-means : Metric space (d, X) α-approximation algorithm: cost less than α*opt k-median :

17 Mapping Core-sets for Capacitated Clustering

18 Capacitated ℓp clustering Problem: Given n points in a metric space, find k centers and assign points to centers, respecting capacities, to minimize ℓp norm of the distance vector. Generalizes balanced k-median, k-means & k-center. Objective is not minimizing cut size (cf. balanced partitioning in the library) Theorem: For any p and k< n, distributed balanced clustering with approx ratio: small constant * best single machine guarantee # rounds: 2 memory: (n/m)2 with m machines Improves [BMVKV 12] and [BEL 13] (Bateni,Bhaskara,Lattanzi,Mirrokni, NIPS 14)

19 Empirical study for distributed clustering Test in terms of scalability and quality of solution Two base instances & subsamples US graph ~30M nodes World graph ~500M nodes Size of seq. inst Increase in OPT US 1/ World 1/ Quality: pessimistic analysis Sublinear running time scaling

20 Submodular maximization Problem: Given k & submodular function f, find set S of size k that maximizes f(s). Some applications Data summarization Feature selection Exemplar clustering Special case: coverage maximization : Given a family of subsets, choose a subfamily of k sets, and maximize cardinality of union. cover various topics/meanings target all kinds of users

21 Submodular maximization Problem: Given k & submodular function f, find set S of size k that maximizes f(s). Some applications Data summarization Feature selection Exemplar clustering Special case: coverage maximization : Given a family of subsets, choose a subfamily of k sets, and maximize cardinality of union. cover various topics/meanings target all kinds of users [IMMM 14] Bad News: No deterministic composable core-set with approx

22 Submodular maximization Problem: Given k & submodular function f, find set S of size k that maximizes f(s). Some applications Data summarization Feature selection Exemplar clustering Special case: coverage maximization : Given a family of subsets, choose a subfamily of k sets, and maximize cardinality of union. cover various topics/meanings target all kinds of users [IMMM 14] Bad News: No deterministic composable core-set with approx Randomization is necessary and useful: Send each set randomly to some machine Build a coreset on each machine by greedy algorithm

23 Randomization to the Rescue: Randomized Core-sets Run GREEDY on each machine Machine 1 Random T1 Random T2 S1 Machine 2 Run GREEDY on selected items to find the final output set S2 Selected Items Input Set Output Set Sm Random Tm Machine m Two rounds of MapReduce

24 Results for Submodular Maximization: MZ (STOC 15) A class of 0.33-approximate randomized composable core-sets of size k for non-monotone submodular maximization. For example, Greedy Algorithm. Hard to go beyond ½ approximation with size k. Impossible to get better than 1-1/e approximate randomized composable core-set of size 4k for monotone f. Results in 0.54-approximate distributed algorithm in two rounds with linear communication complexity. For small-size composable core-sets of k less than k: sqrt{k /k}-approximate randomized composable core-set.

25 Low-Rank Approximation Given (large) matrix A in Rmxn and target rank k << m,n: Optimal solution: k-rank SVD Applications: Dimensionality reduction Signal denoising Compression...

26 Column Subset Selection (CSS) Columns often have important meaning CSS: Low-rank matrix approximation in column space of A n k n k m A A m A[S]

27 DISTGREEDY: GCSS(A,B,k) with L machines B Machine 1 Machine 2 Machine L

28 DISTGREEDY: GCSS(A,B,k) with L machines B Machine 1 Machine 2 Machine L

29 DISTGREEDY: GCSS(A,B,k) with L machines B Machine 1 Machine 2 Machine L Designated machine

30 DISTGREEDY: GCSS(A,B,k) with L machines B Machine 1 Machine 2 Machine L Designated machine

31 DISTGREEDY for column subset selection

32 Empirical result for column subset selection Training accuracy on massive data set (news 20.binary, 15k x 100k matrix) Speedup over 2-phase algorithm in parentheses Interesting experiment: What if we partition more carefully and not randomly? Recent observation: If we treat each machine separately, it does not help much! Random partitioning is good even compared with more careful partitioning.

33 Coverage Problems Problems: Given a set system (n sets and m elements), 1. K-coverage : pick k sets to max. size of union 2. set cover : cover all elements with least number of sets 3. set cover with outliers : cover (1-λ)m elements with least number of sets

34 Coverage Problems Problems: Given a set system (n sets and m elements), 1. K-coverage : pick k sets to max. size of union 2. set cover : cover all elements with least number of sets 3. set cover with outliers : cover (1-λ)m elements with least number of sets Greedy Algorithm: Pick a subset with the maximum marginal coverage,

35 Coverage Problems Problems: Given a set system (n sets and m elements), 1. K-coverage : pick k sets to max. size of union 2. set cover : cover all elements with least number of sets 3. set cover with outliers : cover (1-λ)m elements with least number of sets Greedy Algorithm: Pick a subset with the maximum marginal coverage, 1-1/e-approx. To k-coverage, log n-approximation for set cover... Goal: Achieve good fast approximation with minimum memory footprint Streaming: elements arrive one by one, not sets Distributed: linear communication and memory independent of the size of ground set

36 Submodular Maximization vs. Maximum Coverage Coverage function is a special case of submodular function: f(r) = cardinality of union of family R of subsets

37 Submodular Maximization vs. Maximum Coverage Coverage function is a special case of submodular maximization: f(r) = cardinality of union of family R of subsets So problem solved? [MirrokniZadimoghaddam STOC 15]: Randomized composable core-sets work [Mirzasoleiman et al NIPS 14]: This method works well in Practice!

38 Submodular Maximization vs. Maximum Coverage Coverage function is a special case of submodular maximization: f(r) = cardinality of union of family R of subsets So problem solved? [MirrokniZadimoghaddam STOC 15]: Randomized composable core-sets work [Mirzasoleiman et al NIPS 14]: This method works well in Practice! No. This solution has several issues for coverage problems: It requires expensive oracle access to computing cardinality of union! Distributed Computation: Send whole sets around? Streaming: Handles set arrival model, does not handle element arrival model!

39 Why can t we apply core-sets for submodular functions? Run ALG in each machine Subfamily of subsets T1 Family of subsets Subfamily of subsets Tm Machine 1 T1 Machine 2 T2 S1 Run ALG on selected items to find the final output set S2 Selected Items Output Set Sm Machine m Tm What if the subsets are large? Can we send a sketch of them?

40 Idea: Send a sketch for each set (e.g., sample of elements) Run ALG in each machine Sketch of subsets T1 Family of subsets Sketch of subsets Tm Machine 1 T1 Machine 2 T2 S1 Run ALG on selected items to find the final output set S2 Selected Items Output Set Sm Machine m Tm Question: Does any approximation-preserving sketch work?

41 Approximation-preserving sketching is not sufficient. Idea: Use sketching to define a (1±ε)-approx oracle to cardinality of union function? [BateniEsfandiariMirrokni 16]: Thm 1: A (1±ε)-approx sketch of coverage function May NOT Help Given an (1±ε)-approx oracle to coverage function, we get n0.49 approximation

42 Approximation-preserving sketching is not sufficient. Idea: Use sketching to define a (1±ε)-approx oracle to cardinality of union function? [BateniEsfandiariMirrokni 16]: Thm 1: A (1±ε)-approx sketch of coverage function May NOT Help Given an (1±ε)-approx oracle to coverage function, we get n0.49 approximation Thm 2: With some tricks, MinHash-based sketch + proper sampling WORKS Sample elements not sets (different from previous coreset idea) Correlation between samples (MinHash) Cap degrees of elements in the sketch (reduces memory footprint)

43 Bipartite Graph Formulation for Coverage Problems Bipartite graph G(U, V, E) U: sets V: elements E: membership sets Set cover problem: Pick minimum number of sets that cover all elements. Set cover with outliers problem: Pick minimum number of sets that cover a 1 - fraction of elements. elements Maximum coverage problem: Pick sets that cover maximum number of elements.

44 Sketching Technique sets Construction Dependent sampling: Assign hash values from [0,1) to elements. Remove any element with hash value exceeding. Arbitrarily remove edges to have max-degree for elements. Parameters Sample parameters 1) =is0.6 easy to compute. 2) 2) = 2 be found via a round of MapReduce. can Hash: elements

45 Approach Build graph Sketch construction Core-set method Final greedy Extract results Sketch: sparse subgraph with sufficient information For instance with many sets, parallelize using core sets. Any single-machine greedy algorithm

46 Proof ingredients: 1. Parameters are chosen to produce small sketch (indep. of size of ground set): O(#sets) Challenge: how to choose parameters in distributed or streaming models 2. Any -approximation on the sketch is an + approximation for original instance

47 Summary of Results for Coverage Functions Special case of submodular maximization Problems are NP-hard and APX-hard Greedy algorithm gives best guarantees Good implementations (linear-time) Lazy greedy algorithm Lazier-than-lazy algorithm GREEDY 1) Start with empty solution 2) Until done, (a) find set with best marginal coverage, and (b) add it to tentative solution. Problem: Graph should be stored in RAM Our algorithm: Memory O(#sets) Linear-time Optimal approximation guarantees MapReduce, streaming, etc.

48 Bounds for distributed coverage problems From [BEM 16]: 1) Space indep. of size of sets or ground set, 2) Optimal Approximation Factor, 3) Communication linear in #sets (indep. of their size), 4) small #rounds Previous work: [39]=[CKT 11], [42]=[MZ 15], [19]=[BENW 16], [43]=[MBKK 16]

49 Bounds for streaming coverage problems From [BEM 16]: 1) Space indep. of size of ground set, 2) Optimal Approximation Factor, 3) Edge vs set arrival Previous work:[14]=[cw 15], [22]=[DIMV 14], [24]=[ER 14], [31]=[IMV 15], [49]=[SG 09]

50 Empirical Study Public datasets Social networks Bags of words Contribution graphs Planted instances Very small sketches (0.01 5%) suffice for obtaining good approximations (95+%). Without core sets, can handle in <1h XXXB edges or elements.

51 features Goal: Pick k representative features entities Feature Selection (ongoing) Based on composable core sets k Random clusters Best cluster method Set cover (pairs) ) Pick features that cover all entities 2) Pick features that cover many pairs (or triples, etc.) of entities

52 Summary: Distributed Algorithms for Five Problems Define on a metric space & composable core-sets apply. 1. Diversity Maximization, 2. PODS 14 by IndykMahdianMahabadiM. for Feature Selection in AAAI 17 by AbbasiGhadiriMirrokniZadimoghaddam Capacitated ℓp Clustering, NIPS 14 by BateniBhaskaraLattanziM. Beyond Metric Spaces. Only Randomized partitioning apply Submodular Maximization, STOC 15 by M. Zadimoghaddam Feature Selection (Column Subset Selection), ICML 16 by Alschulter et al. Needs adaptive sampling/sketching techniques 5. Coverage Problems: by BateniEsfandiariM

53 Our team: Google NYC Algorithms Research Team Recently released external team website: research.google.com/teams/nycalg/ Market Algorithms/ Ads Optimization (search & display) Tools: e.g. Clustering common expertise: online allocation problems Large-Scale Graph Mining/Distributed Optimization Infrastructure and Large-Scale Optimization Tools: e.g. Balanced Partitioning

54 THANK YOU

55 Local Search for Diversity Maximization [KDD 13]